TL;DR
This paper introduces a novel method for classifying urban parks into 12 distinct types by fusing urban functional zone (UFZ) data with crowd-sourced Points of Interest (POI) data. By leveraging natural language processing (NLP) to reclassify POIs and then integrating this with UFZ social attributes, the approach achieves a robust classification, offering a more nuanced understanding of urban green spaces than previous methods.
What Problem Does This Paper Address?
Urban parks are crucial for city ecosystems and resident well-being, but existing research often treats urban green spaces as a monolithic entity. This lack of detailed classification hinders effective urban planning, sustainable development analysis, and understanding of how different park types contribute to urban structure and function. Traditional methods relying solely on remote sensing imagery struggle to capture the diverse social functions and attributes of parks, which are better reflected in crowd-sourced data like POIs. The challenge lies in bridging the gap between physical remote sensing data and the rich social information embedded in POIs to accurately identify and categorize the myriad types of urban parks.
Methods
The core of this research is a multi-source data fusion approach for urban park classification. The process involves three main steps:
- POI Reclassification: Original POI data from Gaode Map, which often lacks specific park type information, is reclassified into 12 urban park types. This is achieved through an improved Natural Language Processing (NLP) pipeline that includes:
- Text Similarity Measurement: Using the Jaro-Winkler algorithm to compare POI names with known park types within Urban Functional Zones (UFZs).
- Topic Modeling: Employing a modified TF-IDF model (TF-IWF) to identify latent topics within POI names for unclassified POIs, helping to infer their park type.
- Urban Park Recognition: The reclassified POI data is then integrated with UFZ data. This step uses:
- Attribute Mapping: Determining park types based on the density and types of POIs within UFZ units, with special rules for large "Comprehensive Parks."
- Social Attribute Recognition: Classifying parks into "Monofunctional" and "Multifunctional" categories based on their characteristics and then using UFZ social attributes to refine the classification of "Multifunctional Parks."
- Cartographic Analysis: The final step involves mapping the classified urban parks and conducting comparative analyses between cities.
Datasets
- Urban Functional Zone (UFZ) Data: Obtained from the PKU Urbanscape Essential Dataset, providing high-precision UFZ classifications for Beijing and Hangzhou with a spatial resolution of 2.4m.
- Points of Interest (POI) Data: Collected from Gaode Map, a popular Chinese digital map service. This included 1,052,428 POIs for Beijing and 526,214 for Hangzhou in 2020. Specifically, "tourist attraction" POIs were utilized for park classification.
- OpenStreetMap (OSM) Road Data: Used for precise segmentation and refinement of UFZ boundaries.
Key Insights and Ideas
- POI data is a rich source for social park attributes: While remote sensing captures the physical form, POIs offer crucial insights into the intended use and social function of urban parks, which are vital for accurate classification.
- NLP is key to unlocking POI potential: Automatically reclassifying POI names using text similarity and topic modeling is essential for bridging the gap between raw POI data and specific park typologies.
- Fusing UFZ and POI data offers synergistic benefits: UFZ data provides the spatial context and structure, while POI data adds the social and functional layers, leading to a more comprehensive understanding of urban parks.
- "Monofunctional" vs. "Multifunctional" park classification is a useful heuristic: This distinction helps in applying different classification strategies, leveraging UFZ social attributes more heavily for complex, multifunctional parks.
Key Results
- The proposed method achieved an overall classification accuracy of 82.8% with a Kappa coefficient of 0.74 for urban park mapping in Beijing and Hangzhou.
- Comprehensive Parks were found to be the most prevalent type, with high User's Accuracy (UA) and Producer's Accuracy (PA) (89.8% and 91%, respectively).
- Botanical Gardens, Zoos, and Amusement Parks achieved 100% accuracy, likely due to their distinct characteristics and relatively smaller numbers.
- Historical and Heritage Parks showed lower accuracy (around 66.7% and 52.9% UA, respectively), suggesting that their small size and unique features can lead to misclassification.
- Beijing had a larger total area of urban parks (272.42 km²) compared to Hangzhou (225.92 km²), despite Beijing's smaller overall urban area.
- Beijing had a higher proportion of urban parks (11.02%) than Hangzhou (8.56%).
- Significant differences in park type distribution were observed between Beijing (e.g., high proportion of Comprehensive Parks) and Hangzhou (e.g., high proportion of Forest Parks and Waterfront Parks), reflecting their distinct climates and urban development patterns.
Limitations and Future Work
- UFZ Data Segmentation: The UFZ data used had some issues with over-segmentation and insufficient accuracy, which, despite mitigation efforts using OSM data, still impacted classification efficiency. Future work could benefit from higher-accuracy UFZ data.
- Limited POI Categories: The study primarily used "tourist attraction" POIs, potentially overlooking other relevant data points (e.g., facilities, transportation, distance to city centers) that could further enhance classification.
- Differential Classification Accuracy: The method's accuracy varies across park types, with smaller or more complex types like heritage parks being harder to classify accurately. Future research could focus on more refined POI classification methods.
- In-depth Analysis: Important factors like pedestrian flow, popularity, average park area, vegetation coverage, and accessibility were not considered. Future studies should incorporate these aspects for a more comprehensive urban park analysis.
Summary updated